Elamur Research
☰
home / papers / 2024 / Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models

Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "arXiv — non-exclusive license to distribute" does not allow redistributing the text itself. Original — arxiv.org/pdf/2404.18796

Summary

Grading free-form LLM outputs with one frontier model such as GPT-4 is expensive and can favor the judge's own model family. This Cohere study also shows that such a large judge is often unnecessary. The authors propose PoLL (Panel of LLm evaluators). Several smaller judges from unrelated model families score each output independently, and a voting function merges their verdicts. Their panel uses Command R, Claude-3 Haiku and GPT-3.5. It is tested in three judging regimes across six datasets:

  • reference-based single-hop QA (KILT versions of NQ, TriviaQA, HotpotQA);
  • multi-hop QA (Bamboogle, HotpotQA);
  • pair-wise comparisons on Chatbot Arena Hard.

Agreement is measured against Cohere's in-house human annotators and Chatbot Arena crowd rankings. The panel tracks human decisions more closely than GPT-4 or any single member. It also shows less self-preference and costs over seven times less. Cheaper, more diverse judging setups can therefore be both more reliable and fairer.

Key points

  • Aggregation depends on the task. QA uses max voting over binary correct/incorrect labels, while Arena Hard averages 1–5 scores because three judges rarely form a clear majority.
  • Cohen's κ against humans on KILT (NQ / TQA / HPQA) is 0.763 / 0.906 / 0.867 for PoLL, versus 0.627 / 0.841 / 0.830 for GPT-4. Containment-style EM is weaker on most sets.
  • On Chatbot Arena Hard rankings, PoLL reaches Pearson 0.917 and Kendall Tau 0.778, compared with 0.817 and 0.667 for GPT-4.
  • GPT-4 is sensitive to prompt wording. Its κ on NQ ranges from 0.518 (zero-shot) to 0.725 when told not to overthink, but it still trails Command R, Haiku and PoLL.
  • On multi-hop QA, PoLL's deviation from human accuracy has the lowest spread (std 2.2; GPT-3.5 has 6.1). Individual judges inflate their own family's scores, and GPT-4 ranks a GPT-4 variant 2nd instead of 4th.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Other

Cited by

118 papers of our library cite it (of the 140673 whose bibliographies we read); 2 of them have a page here, 116 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2404.18796 · arXiv 2404.18796

@misc{verga2024replacing,
  title = {{Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models}},
  author = {Verga, Pat and Hofstatter, Sebastian and Althammer, Sophia and Su, Yixuan and Piktus, Aleksandra and Arkhangorodsky, Arkady and Xu, Minjie and White, Naomi and Lewis, Patrick},
  year = {2024},
  doi = {10.48550/arxiv.2404.18796},
  eprint = {2404.18796},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2404.18796}
}

References

31 references, 15 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anthropic. 2024. The claude 3 model family: Opus, sonnet, haiku. find
  2. Bernd Bohnet, Vinh Q Tran, Pat Verga, Roee Aharoni, Daniel Andor, Livio Baldini Soares, Massimiliano Ciaramita, Jacob Eisenstein, Kuzman Ganchev, Jonathan Herzig, et al. 2022. Attributed question answering: Evaluation and modeling for attributed large language models. arXiv preprint arXiv:2212.08037… find our card
  3. Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2020. Measuring massive multitask language understanding. In International Conference on Learning Representations. find our card
  4. Hui Huang, Yingqi Qu, Jing Liu, Muyun Yang, and Tiejun Zhao. 2024. An empirical study of llm-as-a-judge for llm evaluation: Fine-tuned judge models are task-specific classifiers. arXiv preprint arXiv:2403.02839. — [arXiv:2403.02839](https://arxiv.org/abs/2403.02839) find
  5. Mandar Joshi, Eunsol Choi, Daniel Weld, and Luke Zettlemoyer. 2017. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1601–1611, Vancouver… find our card
  6. Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Yu Wu, Sergey Edunov, Danqi Chen, and Wen tau Yih. 2020. Dense passage retrieval for open-domain question answering. ArXiv, abs/2004.04906. find our card
  7. Maurice G Kendall. 1938. A new measure of rank correlation. Biometrika, 30(1/2):81–93. find
  8. Tom Kocmi and Christian Federmann. 2023a. Large language models are state-of-the-art evaluators of translation quality. In Proceedings of the 24th Annual Conference of the European Association for Machine Translation, pages 193–203, Tampere, Finland. European Association for Machine Translation. find
  9. Tom Kocmi and Christian Federmann. 2023b. Large language models are state-of-the-art evaluators of translation quality. arXiv preprint arXiv:2302.14520. find
  10. Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural… find our card
  11. Ruosen Li, Teerth Patel, and Xinya Du. 2023. Prd: Peer rank and discussion improve large language model based evaluations. arXiv preprint arXiv:2307.02762. find
  12. Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Banghua Zhu, Joseph E. Gonzalez, and Ion Stoica. 2024a. From live data to high-quality benchmarks: The arena-hard pipeline. find
  13. Zhen Li, Xiaohan Xu, Tao Shen, Can Xu, Jia-Chen Gu, and Chongyang Tao. 2024b. Leveraging large language models for nlg evaluation: A survey. arXiv e-prints, pages arXiv–2401. find
  14. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. find
  15. Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12:157–173. find our card
  16. OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Ja… find our card
  17. Arjun Panickssery, Samuel R. Bowman, and Shi Feng. 2024. Llm evaluators recognize and favor their own generations. arXiv e-prints, pages arXiv–2401. find
  18. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311–318. find our card
  19. Karl Pearson. 1895. Vii. note on regression and inheritance in the case of two parents. proceedings of the royal society of London, 58(347-352):240–242. find
  20. Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, et al. 2021. Kilt: a benchmark for knowledge intensive language tasks. In Proceedings of the 2021 Conference of the North American Chapter of th… find our card
  21. Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. 2023. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 5687–5711. find
  22. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383–2392. find our card
  23. Thibault Sellam, Dipanjan Das, and Ankur Parikh. 2020. BLEURT: Learning robust metrics for text generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7881–7892, Online. Association for Computational Linguistics. find
  24. Chenhui Shen, Liying Cheng, Xuan-Phi Nguyen, Yang You, and Lidong Bing. 2023. Large language models are not yet human-level evaluators for abstractive summarization. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 4215–4233, Singapore. Association for Computational Li… find
  25. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36. find our card
  26. Ellen M. Voorhees. 1998. Variations in relevance judgments and the measurement of retrieval effectiveness. In Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’98, page 315–323, New York, NY, USA. Association for Computing… find
  27. Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. Improving text embeddings with large language models. ArXiv, abs/2401.00368. find our card
  28. Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369… find our card
  29. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. arXiv preprint arXiv:2210.03629. find our card
  30. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36. find
  31. Lianghui Zhu, Xinggang Wang, and Xinlong Wang. 2023. Judgelm: Fine-tuned large language models are scalable judges. arXiv preprint arXiv:2310.17631. find our card